refactored StereoBeliefPropagation

modules/gpustereo/include/opencv2/gpustereo.hpp

@@ -65,61 +65,55 @@ public:
 
 CV_EXPORTS Ptr<gpu::StereoBM> createStereoBM(int numDisparities = 64, int blockSize = 19);
 
+/////////////////////////////////////////
+// StereoBeliefPropagation
 
-
-// "Efficient Belief Propagation for Early Vision"
-// P.Felzenszwalb
-class CV_EXPORTS StereoBeliefPropagation
+//! "Efficient Belief Propagation for Early Vision" P.Felzenszwalb
+class CV_EXPORTS StereoBeliefPropagation : public cv::StereoMatcher
 {
 public:
-    enum { DEFAULT_NDISP  = 64 };
-    enum { DEFAULT_ITERS  = 5  };
-    enum { DEFAULT_LEVELS = 5  };
-
-    static void estimateRecommendedParams(int width, int height, int& ndisp, int& iters, int& levels);
-
-    //! the default constructor
-    explicit StereoBeliefPropagation(int ndisp  = DEFAULT_NDISP,
-                                     int iters  = DEFAULT_ITERS,
-                                     int levels = DEFAULT_LEVELS,
-                                     int msg_type = CV_32F);
-
-    //! the full constructor taking the number of disparities, number of BP iterations on each level,
-    //! number of levels, truncation of data cost, data weight,
-    //! truncation of discontinuity cost and discontinuity single jump
-    //! DataTerm = data_weight * min(fabs(I2-I1), max_data_term)
-    //! DiscTerm = min(disc_single_jump * fabs(f1-f2), max_disc_term)
-    //! please see paper for more details
-    StereoBeliefPropagation(int ndisp, int iters, int levels,
-        float max_data_term, float data_weight,
-        float max_disc_term, float disc_single_jump,
-        int msg_type = CV_32F);
-
-    //! the stereo correspondence operator. Finds the disparity for the specified rectified stereo pair,
-    //! if disparity is empty output type will be CV_16S else output type will be disparity.type().
-    void operator()(const GpuMat& left, const GpuMat& right, GpuMat& disparity, Stream& stream = Stream::Null());
+    using cv::StereoMatcher::compute;
 
+    virtual void compute(InputArray left, InputArray right, OutputArray disparity, Stream& stream) = 0;
 
     //! version for user specified data term
-    void operator()(const GpuMat& data, GpuMat& disparity, Stream& stream = Stream::Null());
+    virtual void compute(InputArray data, OutputArray disparity, Stream& stream = Stream::Null()) = 0;
 
-    int ndisp;
+    //! number of BP iterations on each level
+    virtual int getNumIters() const = 0;
+    virtual void setNumIters(int iters) = 0;
 
-    int iters;
-    int levels;
+    //! number of levels
+    virtual int getNumLevels() const = 0;
+    virtual void setNumLevels(int levels) = 0;
 
-    float max_data_term;
-    float data_weight;
-    float max_disc_term;
-    float disc_single_jump;
+    //! truncation of data cost
+    virtual double getMaxDataTerm() const = 0;
+    virtual void setMaxDataTerm(double max_data_term) = 0;
 
-    int msg_type;
-private:
-    GpuMat u, d, l, r, u2, d2, l2, r2;
-    std::vector<GpuMat> datas;
-    GpuMat out;
+    //! data weight
+    virtual double getDataWeight() const = 0;
+    virtual void setDataWeight(double data_weight) = 0;
+
+    //! truncation of discontinuity cost
+    virtual double getMaxDiscTerm() const = 0;
+    virtual void setMaxDiscTerm(double max_disc_term) = 0;
+
+    //! discontinuity single jump
+    virtual double getDiscSingleJump() const = 0;
+    virtual void setDiscSingleJump(double disc_single_jump) = 0;
+
+    virtual int getMsgType() const = 0;
+    virtual void setMsgType(int msg_type) = 0;
+
+    static void estimateRecommendedParams(int width, int height, int& ndisp, int& iters, int& levels);
 };
 
+CV_EXPORTS Ptr<gpu::StereoBeliefPropagation>
+    createStereoBeliefPropagation(int ndisp = 64, int iters = 5, int levels = 5, int msg_type = CV_32F);
+
+
+
 // "A Constant-Space Belief Propagation Algorithm for Stereo Matching"
 // Qingxiong Yang, Liang Wang, Narendra Ahuja
 // http://vision.ai.uiuc.edu/~qyang6/

modules/gpustereo/perf/perf_stereo.cpp

@@ -107,13 +107,13 @@ PERF_TEST_P(ImagePair, StereoBeliefPropagation,
 
     if (PERF_RUN_GPU())
     {
-        cv::gpu::StereoBeliefPropagation d_bp(ndisp);
+        cv::Ptr<cv::gpu::StereoBeliefPropagation> d_bp = cv::gpu::createStereoBeliefPropagation(ndisp);
 
         const cv::gpu::GpuMat d_imgLeft(imgLeft);
         const cv::gpu::GpuMat d_imgRight(imgRight);
         cv::gpu::GpuMat dst;
 
-        TEST_CYCLE() d_bp(d_imgLeft, d_imgRight, dst);
+        TEST_CYCLE() d_bp->compute(d_imgLeft, d_imgRight, dst);
 
         GPU_SANITY_CHECK(dst);
     }

modules/gpustereo/src/stereobp.cpp

@@ -49,12 +49,7 @@ using namespace cv::gpu;
 
 void cv::gpu::StereoBeliefPropagation::estimateRecommendedParams(int, int, int&, int&, int&) { throw_no_cuda(); }
 
-cv::gpu::StereoBeliefPropagation::StereoBeliefPropagation(int, int, int, int) { throw_no_cuda(); }
-cv::gpu::StereoBeliefPropagation::StereoBeliefPropagation(int, int, int, float, float, float, float, int) { throw_no_cuda(); }
-
-void cv::gpu::StereoBeliefPropagation::operator()(const GpuMat&, const GpuMat&, GpuMat&, Stream&) { throw_no_cuda(); }
-
-void cv::gpu::StereoBeliefPropagation::operator()(const GpuMat&, GpuMat&, Stream&) { throw_no_cuda(); }
+Ptr<gpu::StereoBeliefPropagation> cv::gpu::createStereoBeliefPropagation(int, int, int, int) { throw_no_cuda(); return Ptr<gpu::StereoBeliefPropagation>(); }
 
 #else /* !defined (HAVE_CUDA) */
 
@@ -78,14 +73,295 @@ namespace cv { namespace gpu { namespace cudev
     }
 }}}
 
-using namespace ::cv::gpu::cudev::stereobp;
-
 namespace
 {
+    class StereoBPImpl : public gpu::StereoBeliefPropagation
+    {
+    public:
+        StereoBPImpl(int ndisp, int iters, int levels, int msg_type);
+
+        void compute(InputArray left, InputArray right, OutputArray disparity);
+        void compute(InputArray left, InputArray right, OutputArray disparity, Stream& stream);
+        void compute(InputArray data, OutputArray disparity, Stream& stream);
+
+        int getMinDisparity() const { return 0; }
+        void setMinDisparity(int /*minDisparity*/) {}
+
+        int getNumDisparities() const { return ndisp_; }
+        void setNumDisparities(int numDisparities) { ndisp_ = numDisparities; }
+
+        int getBlockSize() const { return 0; }
+        void setBlockSize(int /*blockSize*/) {}
+
+        int getSpeckleWindowSize() const { return 0; }
+        void setSpeckleWindowSize(int /*speckleWindowSize*/) {}
+
+        int getSpeckleRange() const { return 0; }
+        void setSpeckleRange(int /*speckleRange*/) {}
+
+        int getDisp12MaxDiff() const { return 0; }
+        void setDisp12MaxDiff(int /*disp12MaxDiff*/) {}
+
+        int getNumIters() const { return iters_; }
+        void setNumIters(int iters) { iters_ = iters; }
+
+        int getNumLevels() const { return levels_; }
+        void setNumLevels(int levels) { levels_ = levels; }
+
+        double getMaxDataTerm() const { return max_data_term_; }
+        void setMaxDataTerm(double max_data_term) { max_data_term_ = (float) max_data_term; }
+
+        double getDataWeight() const { return data_weight_; }
+        void setDataWeight(double data_weight) { data_weight_ = (float) data_weight; }
+
+        double getMaxDiscTerm() const { return max_disc_term_; }
+        void setMaxDiscTerm(double max_disc_term) { max_disc_term_ = (float) max_disc_term; }
+
+        double getDiscSingleJump() const { return disc_single_jump_; }
+        void setDiscSingleJump(double disc_single_jump) { disc_single_jump_ = (float) disc_single_jump; }
+
+        int getMsgType() const { return msg_type_; }
+        void setMsgType(int msg_type) { msg_type_ = msg_type; }
+
+    private:
+        void init(Stream& stream);
+        void calcBP(OutputArray disp, Stream& stream);
+
+        int ndisp_;
+        int iters_;
+        int levels_;
+        float max_data_term_;
+        float data_weight_;
+        float max_disc_term_;
+        float disc_single_jump_;
+        int msg_type_;
+
+        float scale_;
+        int rows_, cols_;
+        std::vector<int> cols_all_, rows_all_;
+        GpuMat u_, d_, l_, r_, u2_, d2_, l2_, r2_;
+        std::vector<GpuMat> datas_;
+        GpuMat outBuf_;
+    };
+
     const float DEFAULT_MAX_DATA_TERM = 10.0f;
     const float DEFAULT_DATA_WEIGHT = 0.07f;
     const float DEFAULT_MAX_DISC_TERM = 1.7f;
     const float DEFAULT_DISC_SINGLE_JUMP = 1.0f;
+
+    StereoBPImpl::StereoBPImpl(int ndisp, int iters, int levels, int msg_type) :
+        ndisp_(ndisp), iters_(iters), levels_(levels),
+        max_data_term_(DEFAULT_MAX_DATA_TERM), data_weight_(DEFAULT_DATA_WEIGHT),
+        max_disc_term_(DEFAULT_MAX_DISC_TERM), disc_single_jump_(DEFAULT_DISC_SINGLE_JUMP),
+        msg_type_(msg_type)
+    {
+    }
+
+    void StereoBPImpl::compute(InputArray left, InputArray right, OutputArray disparity)
+    {
+        compute(left, right, disparity, Stream::Null());
+    }
+
+    void StereoBPImpl::compute(InputArray _left, InputArray _right, OutputArray disparity, Stream& stream)
+    {
+        using namespace cv::gpu::cudev::stereobp;
+
+        typedef void (*comp_data_t)(const PtrStepSzb& left, const PtrStepSzb& right, const PtrStepSzb& data, cudaStream_t stream);
+        static const comp_data_t comp_data_callers[2][5] =
+        {
+            {0, comp_data_gpu<unsigned char, short>, 0, comp_data_gpu<uchar3, short>, comp_data_gpu<uchar4, short>},
+            {0, comp_data_gpu<unsigned char, float>, 0, comp_data_gpu<uchar3, float>, comp_data_gpu<uchar4, float>}
+        };
+
+        scale_ = msg_type_ == CV_32F ? 1.0f : 10.0f;
+
+        CV_Assert( 0 < ndisp_ && 0 < iters_ && 0 < levels_ );
+        CV_Assert( msg_type_ == CV_32F || msg_type_ == CV_16S );
+        CV_Assert( msg_type_ == CV_32F || (1 << (levels_ - 1)) * scale_ * max_data_term_ < std::numeric_limits<short>::max() );
+
+        GpuMat left = _left.getGpuMat();
+        GpuMat right = _right.getGpuMat();
+
+        CV_Assert( left.type() == CV_8UC1 || left.type() == CV_8UC3 || left.type() == CV_8UC4 );
+        CV_Assert( left.size() == right.size() && left.type() == right.type() );
+
+        rows_ = left.rows;
+        cols_ = left.cols;
+
+        const int divisor = (int) pow(2.f, levels_ - 1.0f);
+        const int lowest_cols = cols_ / divisor;
+        const int lowest_rows = rows_ / divisor;
+        const int min_image_dim_size = 2;
+        CV_Assert( std::min(lowest_cols, lowest_rows) > min_image_dim_size );
+
+        init(stream);
+
+        datas_[0].create(rows_ * ndisp_, cols_, msg_type_);
+
+        comp_data_callers[msg_type_ == CV_32F][left.channels()](left, right, datas_[0], StreamAccessor::getStream(stream));
+
+        calcBP(disparity, stream);
+    }
+
+    void StereoBPImpl::compute(InputArray _data, OutputArray disparity, Stream& stream)
+    {
+        scale_ = msg_type_ == CV_32F ? 1.0f : 10.0f;
+
+        CV_Assert( 0 < ndisp_ && 0 < iters_ && 0 < levels_ );
+        CV_Assert( msg_type_ == CV_32F || msg_type_ == CV_16S );
+        CV_Assert( msg_type_ == CV_32F || (1 << (levels_ - 1)) * scale_ * max_data_term_ < std::numeric_limits<short>::max() );
+
+        GpuMat data = _data.getGpuMat();
+
+        CV_Assert( (data.type() == msg_type_) && (data.rows % ndisp_ == 0) );
+
+        rows_ = data.rows / ndisp_;
+        cols_ = data.cols;
+
+        const int divisor = (int) pow(2.f, levels_ - 1.0f);
+        const int lowest_cols = cols_ / divisor;
+        const int lowest_rows = rows_ / divisor;
+        const int min_image_dim_size = 2;
+        CV_Assert( std::min(lowest_cols, lowest_rows) > min_image_dim_size );
+
+        init(stream);
+
+        data.copyTo(datas_[0], stream);
+
+        calcBP(disparity, stream);
+    }
+
+    void StereoBPImpl::init(Stream& stream)
+    {
+        using namespace cv::gpu::cudev::stereobp;
+
+        u_.create(rows_ * ndisp_, cols_, msg_type_);
+        d_.create(rows_ * ndisp_, cols_, msg_type_);
+        l_.create(rows_ * ndisp_, cols_, msg_type_);
+        r_.create(rows_ * ndisp_, cols_, msg_type_);
+
+        if (levels_ & 1)
+        {
+            //can clear less area
+            u_.setTo(0, stream);
+            d_.setTo(0, stream);
+            l_.setTo(0, stream);
+            r_.setTo(0, stream);
+        }
+
+        if (levels_ > 1)
+        {
+            int less_rows = (rows_ + 1) / 2;
+            int less_cols = (cols_ + 1) / 2;
+
+            u2_.create(less_rows * ndisp_, less_cols, msg_type_);
+            d2_.create(less_rows * ndisp_, less_cols, msg_type_);
+            l2_.create(less_rows * ndisp_, less_cols, msg_type_);
+            r2_.create(less_rows * ndisp_, less_cols, msg_type_);
+
+            if ((levels_ & 1) == 0)
+            {
+                u2_.setTo(0, stream);
+                d2_.setTo(0, stream);
+                l2_.setTo(0, stream);
+                r2_.setTo(0, stream);
+            }
+        }
+
+        load_constants(ndisp_, max_data_term_, scale_ * data_weight_, scale_ * max_disc_term_, scale_ * disc_single_jump_);
+
+        datas_.resize(levels_);
+
+        cols_all_.resize(levels_);
+        rows_all_.resize(levels_);
+
+        cols_all_[0] = cols_;
+        rows_all_[0] = rows_;
+    }
+
+    void StereoBPImpl::calcBP(OutputArray disp, Stream& _stream)
+    {
+        using namespace cv::gpu::cudev::stereobp;
+
+        typedef void (*data_step_down_t)(int dst_cols, int dst_rows, int src_rows, const PtrStepSzb& src, const PtrStepSzb& dst, cudaStream_t stream);
+        static const data_step_down_t data_step_down_callers[2] =
+        {
+            data_step_down_gpu<short>, data_step_down_gpu<float>
+        };
+
+        typedef void (*level_up_messages_t)(int dst_idx, int dst_cols, int dst_rows, int src_rows, PtrStepSzb* mus, PtrStepSzb* mds, PtrStepSzb* mls, PtrStepSzb* mrs, cudaStream_t stream);
+        static const level_up_messages_t level_up_messages_callers[2] =
+        {
+            level_up_messages_gpu<short>, level_up_messages_gpu<float>
+        };
+
+        typedef void (*calc_all_iterations_t)(int cols, int rows, int iters, const PtrStepSzb& u, const PtrStepSzb& d, const PtrStepSzb& l, const PtrStepSzb& r, const PtrStepSzb& data, cudaStream_t stream);
+        static const calc_all_iterations_t calc_all_iterations_callers[2] =
+        {
+            calc_all_iterations_gpu<short>, calc_all_iterations_gpu<float>
+        };
+
+        typedef void (*output_t)(const PtrStepSzb& u, const PtrStepSzb& d, const PtrStepSzb& l, const PtrStepSzb& r, const PtrStepSzb& data, const PtrStepSz<short>& disp, cudaStream_t stream);
+        static const output_t output_callers[2] =
+        {
+            output_gpu<short>, output_gpu<float>
+        };
+
+        const int funcIdx = msg_type_ == CV_32F;
+
+        cudaStream_t stream = StreamAccessor::getStream(_stream);
+
+        for (int i = 1; i < levels_; ++i)
+        {
+            cols_all_[i] = (cols_all_[i-1] + 1) / 2;
+            rows_all_[i] = (rows_all_[i-1] + 1) / 2;
+
+            datas_[i].create(rows_all_[i] * ndisp_, cols_all_[i], msg_type_);
+
+            data_step_down_callers[funcIdx](cols_all_[i], rows_all_[i], rows_all_[i-1], datas_[i-1], datas_[i], stream);
+        }
+
+        PtrStepSzb mus[] = {u_, u2_};
+        PtrStepSzb mds[] = {d_, d2_};
+        PtrStepSzb mrs[] = {r_, r2_};
+        PtrStepSzb mls[] = {l_, l2_};
+
+        int mem_idx = (levels_ & 1) ? 0 : 1;
+
+        for (int i = levels_ - 1; i >= 0; --i)
+        {
+            // for lower level we have already computed messages by setting to zero
+            if (i != levels_ - 1)
+                level_up_messages_callers[funcIdx](mem_idx, cols_all_[i], rows_all_[i], rows_all_[i+1], mus, mds, mls, mrs, stream);
+
+            calc_all_iterations_callers[funcIdx](cols_all_[i], rows_all_[i], iters_, mus[mem_idx], mds[mem_idx], mls[mem_idx], mrs[mem_idx], datas_[i], stream);
+
+            mem_idx = (mem_idx + 1) & 1;
+        }
+
+        const int dtype = disp.fixedType() ? disp.type() : CV_16SC1;
+
+        disp.create(rows_, cols_, dtype);
+        GpuMat out = disp.getGpuMat();
+
+        if (dtype != CV_16SC1)
+        {
+            outBuf_.create(rows_, cols_, CV_16SC1);
+            out = outBuf_;
+        }
+
+        out.setTo(0, _stream);
+
+        output_callers[funcIdx](u_, d_, l_, r_, datas_.front(), out, stream);
+
+        if (dtype != CV_16SC1)
+            out.convertTo(disp, dtype, _stream);
+    }
+}
+
+Ptr<gpu::StereoBeliefPropagation> cv::gpu::createStereoBeliefPropagation(int ndisp, int iters, int levels, int msg_type)
+{
+    return new StereoBPImpl(ndisp, iters, levels, msg_type);
 }
 
 void cv::gpu::StereoBeliefPropagation::estimateRecommendedParams(int width, int height, int& ndisp, int& iters, int& levels)
@@ -101,240 +377,4 @@ void cv::gpu::StereoBeliefPropagation::estimateRecommendedParams(int width, int
     if (levels == 0) levels++;
 }
 
-cv::gpu::StereoBeliefPropagation::StereoBeliefPropagation(int ndisp_, int iters_, int levels_, int msg_type_)
-    : ndisp(ndisp_), iters(iters_), levels(levels_),
-      max_data_term(DEFAULT_MAX_DATA_TERM), data_weight(DEFAULT_DATA_WEIGHT),
-      max_disc_term(DEFAULT_MAX_DISC_TERM), disc_single_jump(DEFAULT_DISC_SINGLE_JUMP),
-      msg_type(msg_type_), datas(levels_)
-{
-}
-
-cv::gpu::StereoBeliefPropagation::StereoBeliefPropagation(int ndisp_, int iters_, int levels_, float max_data_term_, float data_weight_, float max_disc_term_, float disc_single_jump_, int msg_type_)
-    : ndisp(ndisp_), iters(iters_), levels(levels_),
-      max_data_term(max_data_term_), data_weight(data_weight_),
-      max_disc_term(max_disc_term_), disc_single_jump(disc_single_jump_),
-      msg_type(msg_type_), datas(levels_)
-{
-}
-
-namespace
-{
-    class StereoBeliefPropagationImpl
-    {
-    public:
-        StereoBeliefPropagationImpl(StereoBeliefPropagation& rthis_,
-                                    GpuMat& u_, GpuMat& d_, GpuMat& l_, GpuMat& r_,
-                                    GpuMat& u2_, GpuMat& d2_, GpuMat& l2_, GpuMat& r2_,
-                                    std::vector<GpuMat>& datas_, GpuMat& out_)
-            : rthis(rthis_), u(u_), d(d_), l(l_), r(r_), u2(u2_), d2(d2_), l2(l2_), r2(r2_), datas(datas_), out(out_),
-              zero(Scalar::all(0)), scale(rthis_.msg_type == CV_32F ? 1.0f : 10.0f)
-        {
-            CV_Assert(0 < rthis.ndisp && 0 < rthis.iters && 0 < rthis.levels);
-            CV_Assert(rthis.msg_type == CV_32F || rthis.msg_type == CV_16S);
-            CV_Assert(rthis.msg_type == CV_32F || (1 << (rthis.levels - 1)) * scale * rthis.max_data_term < std::numeric_limits<short>::max());
-        }
-
-        void operator()(const GpuMat& left, const GpuMat& right, GpuMat& disp, Stream& stream)
-        {
-            typedef void (*comp_data_t)(const PtrStepSzb& left, const PtrStepSzb& right, const PtrStepSzb& data, cudaStream_t stream);
-            static const comp_data_t comp_data_callers[2][5] =
-            {
-                {0, comp_data_gpu<unsigned char, short>, 0, comp_data_gpu<uchar3, short>, comp_data_gpu<uchar4, short>},
-                {0, comp_data_gpu<unsigned char, float>, 0, comp_data_gpu<uchar3, float>, comp_data_gpu<uchar4, float>}
-            };
-
-            CV_Assert(left.size() == right.size() && left.type() == right.type());
-            CV_Assert(left.type() == CV_8UC1 || left.type() == CV_8UC3 || left.type() == CV_8UC4);
-
-            rows = left.rows;
-            cols = left.cols;
-
-            int divisor = (int)pow(2.f, rthis.levels - 1.0f);
-            int lowest_cols = cols / divisor;
-            int lowest_rows = rows / divisor;
-            const int min_image_dim_size = 2;
-            CV_Assert(std::min(lowest_cols, lowest_rows) > min_image_dim_size);
-
-            init(stream);
-
-            datas[0].create(rows * rthis.ndisp, cols, rthis.msg_type);
-
-            comp_data_callers[rthis.msg_type == CV_32F][left.channels()](left, right, datas[0], StreamAccessor::getStream(stream));
-
-            calcBP(disp, stream);
-        }
-
-        void operator()(const GpuMat& data, GpuMat& disp, Stream& stream)
-        {
-            CV_Assert((data.type() == rthis.msg_type) && (data.rows % rthis.ndisp == 0));
-
-            rows = data.rows / rthis.ndisp;
-            cols = data.cols;
-
-            int divisor = (int)pow(2.f, rthis.levels - 1.0f);
-            int lowest_cols = cols / divisor;
-            int lowest_rows = rows / divisor;
-            const int min_image_dim_size = 2;
-            CV_Assert(std::min(lowest_cols, lowest_rows) > min_image_dim_size);
-
-            init(stream);
-
-            datas[0] = data;
-
-            calcBP(disp, stream);
-        }
-    private:
-        void init(Stream& stream)
-        {
-            u.create(rows * rthis.ndisp, cols, rthis.msg_type);
-            d.create(rows * rthis.ndisp, cols, rthis.msg_type);
-            l.create(rows * rthis.ndisp, cols, rthis.msg_type);
-            r.create(rows * rthis.ndisp, cols, rthis.msg_type);
-
-            if (rthis.levels & 1)
-            {
-                //can clear less area
-                u.setTo(zero, stream);
-                d.setTo(zero, stream);
-                l.setTo(zero, stream);
-                r.setTo(zero, stream);
-            }
-
-            if (rthis.levels > 1)
-            {
-                int less_rows = (rows + 1) / 2;
-                int less_cols = (cols + 1) / 2;
-
-                u2.create(less_rows * rthis.ndisp, less_cols, rthis.msg_type);
-                d2.create(less_rows * rthis.ndisp, less_cols, rthis.msg_type);
-                l2.create(less_rows * rthis.ndisp, less_cols, rthis.msg_type);
-                r2.create(less_rows * rthis.ndisp, less_cols, rthis.msg_type);
-
-                if ((rthis.levels & 1) == 0)
-                {
-                    u2.setTo(zero, stream);
-                    d2.setTo(zero, stream);
-                    l2.setTo(zero, stream);
-                    r2.setTo(zero, stream);
-                }
-            }
-
-            load_constants(rthis.ndisp, rthis.max_data_term, scale * rthis.data_weight, scale * rthis.max_disc_term, scale * rthis.disc_single_jump);
-
-            datas.resize(rthis.levels);
-
-            cols_all.resize(rthis.levels);
-            rows_all.resize(rthis.levels);
-
-            cols_all[0] = cols;
-            rows_all[0] = rows;
-        }
-
-        void calcBP(GpuMat& disp, Stream& stream)
-        {
-            typedef void (*data_step_down_t)(int dst_cols, int dst_rows, int src_rows, const PtrStepSzb& src, const PtrStepSzb& dst, cudaStream_t stream);
-            static const data_step_down_t data_step_down_callers[2] =
-            {
-                data_step_down_gpu<short>, data_step_down_gpu<float>
-            };
-
-            typedef void (*level_up_messages_t)(int dst_idx, int dst_cols, int dst_rows, int src_rows, PtrStepSzb* mus, PtrStepSzb* mds, PtrStepSzb* mls, PtrStepSzb* mrs, cudaStream_t stream);
-            static const level_up_messages_t level_up_messages_callers[2] =
-            {
-                level_up_messages_gpu<short>, level_up_messages_gpu<float>
-            };
-
-            typedef void (*calc_all_iterations_t)(int cols, int rows, int iters, const PtrStepSzb& u, const PtrStepSzb& d, const PtrStepSzb& l, const PtrStepSzb& r, const PtrStepSzb& data, cudaStream_t stream);
-            static const calc_all_iterations_t calc_all_iterations_callers[2] =
-            {
-                calc_all_iterations_gpu<short>, calc_all_iterations_gpu<float>
-            };
-
-            typedef void (*output_t)(const PtrStepSzb& u, const PtrStepSzb& d, const PtrStepSzb& l, const PtrStepSzb& r, const PtrStepSzb& data, const PtrStepSz<short>& disp, cudaStream_t stream);
-            static const output_t output_callers[2] =
-            {
-                output_gpu<short>, output_gpu<float>
-            };
-
-            const int funcIdx = rthis.msg_type == CV_32F;
-
-            cudaStream_t cudaStream = StreamAccessor::getStream(stream);
-
-            for (int i = 1; i < rthis.levels; ++i)
-            {
-                cols_all[i] = (cols_all[i-1] + 1) / 2;
-                rows_all[i] = (rows_all[i-1] + 1) / 2;
-
-                datas[i].create(rows_all[i] * rthis.ndisp, cols_all[i], rthis.msg_type);
-
-                data_step_down_callers[funcIdx](cols_all[i], rows_all[i], rows_all[i-1], datas[i-1], datas[i], cudaStream);
-            }
-
-            PtrStepSzb mus[] = {u, u2};
-            PtrStepSzb mds[] = {d, d2};
-            PtrStepSzb mrs[] = {r, r2};
-            PtrStepSzb mls[] = {l, l2};
-
-            int mem_idx = (rthis.levels & 1) ? 0 : 1;
-
-            for (int i = rthis.levels - 1; i >= 0; --i)
-            {
-                // for lower level we have already computed messages by setting to zero
-                if (i != rthis.levels - 1)
-                    level_up_messages_callers[funcIdx](mem_idx, cols_all[i], rows_all[i], rows_all[i+1], mus, mds, mls, mrs, cudaStream);
-
-                calc_all_iterations_callers[funcIdx](cols_all[i], rows_all[i], rthis.iters, mus[mem_idx], mds[mem_idx], mls[mem_idx], mrs[mem_idx], datas[i], cudaStream);
-
-                mem_idx = (mem_idx + 1) & 1;
-            }
-
-            if (disp.empty())
-                disp.create(rows, cols, CV_16S);
-
-            out = ((disp.type() == CV_16S) ? disp : (out.create(rows, cols, CV_16S), out));
-
-            out.setTo(zero, stream);
-
-            output_callers[funcIdx](u, d, l, r, datas.front(), out, cudaStream);
-
-            if (disp.type() != CV_16S)
-                out.convertTo(disp, disp.type(), stream);
-        }
-
-        StereoBeliefPropagation& rthis;
-
-        GpuMat& u;
-        GpuMat& d;
-        GpuMat& l;
-        GpuMat& r;
-
-        GpuMat& u2;
-        GpuMat& d2;
-        GpuMat& l2;
-        GpuMat& r2;
-
-        std::vector<GpuMat>& datas;
-        GpuMat& out;
-
-        const Scalar zero;
-        const float scale;
-
-        int rows, cols;
-
-        std::vector<int> cols_all, rows_all;
-    };
-}
-
-void cv::gpu::StereoBeliefPropagation::operator()(const GpuMat& left, const GpuMat& right, GpuMat& disp, Stream& stream)
-{
-    StereoBeliefPropagationImpl impl(*this, u, d, l, r, u2, d2, l2, r2, datas, out);
-    impl(left, right, disp, stream);
-}
-
-void cv::gpu::StereoBeliefPropagation::operator()(const GpuMat& data, GpuMat& disp, Stream& stream)
-{
-    StereoBeliefPropagationImpl impl(*this, u, d, l, r, u2, d2, l2, r2, datas, out);
-    impl(data, disp, stream);
-}
-
 #endif /* !defined (HAVE_CUDA) */

modules/gpustereo/test/test_stereo.cpp

@@ -106,10 +106,15 @@ GPU_TEST_P(StereoBeliefPropagation, Regression)
     ASSERT_FALSE(right_image.empty());
     ASSERT_FALSE(disp_gold.empty());
 
-    cv::gpu::StereoBeliefPropagation bp(64, 8, 2, 25, 0.1f, 15, 1, CV_16S);
+    cv::Ptr<cv::gpu::StereoBeliefPropagation> bp = cv::gpu::createStereoBeliefPropagation(64, 8, 2, CV_16S);
+    bp->setMaxDataTerm(25.0);
+    bp->setDataWeight(0.1);
+    bp->setMaxDiscTerm(15.0);
+    bp->setDiscSingleJump(1.0);
+
     cv::gpu::GpuMat disp;
 
-    bp(loadMat(left_image), loadMat(right_image), disp);
+    bp->compute(loadMat(left_image), loadMat(right_image), disp);
 
     cv::Mat h_disp(disp);
     h_disp.convertTo(h_disp, disp_gold.depth());